Auto-Association and Novelty Detection: Truths and Myths?

Abstract

In many damage detection approaches it is often impossible, or very difficult, to obtain true measurements for all possible damage classes, especially in complex structures. In fact, data that is collected during any damaged state of a structure is usually very rare. This is the reason that unsupervised learning is critical. For databased approaches to structural health monitoring the premise of novelty detection techniques is to seek the answer to a simple question; given a newly presented measurement from the structure, does one believe it to have come from the structure in its undamaged state? Auto-associative neural networks consisting of five layers have been used as an advanced method for novelty detection in the past. In this study an analysis is performed in order to demonstrate the ability of nonlinear autoassociators with three layers for multimodal classification problems and novelty detection. Also, another critical issue, that of generalisation in the context of autoassociators is investigated and discussed. Is generalisation needed for novelty detection when neural networks are used or not?